Claude Projects vs Custom GPTs vs a Real Agent: What to Use When

Last month a client paid a developer $9,000 to build a “custom AI agent” for proposal drafting. It was a Claude Project with three files in it, wrapped in a web app nobody asked for. The actual work, the part that made the proposals good, was the writing samples and the pricing sheet he uploaded. He could have done that himself in an afternoon for the cost of a Claude subscription.

This is the most common expensive mistake I see right now. People reach for “an agent” because it sounds like the serious answer, when a Claude Project or a Custom GPT would have solved the whole thing. The three get conflated constantly, and the gap between them is real money and real maintenance.

So here is the framework I actually use when someone asks me which one they need. It comes down to three variables: memory, repeatability, and who else touches it.

What’s the actual difference

Strip the marketing and these three things are not competitors. They sit at different points on a single line, from “smart chat with context” to “software that runs without you.”

A Claude Project (or a Custom GPT, they are close cousins) is a configured chat workspace. You give it instructions, drop in reference files, and it remembers all of that for every conversation inside that project. You are still in the loop. You type, it answers, you steer.

A Custom GPT is the same idea inside ChatGPT, with one extra trick: it can call out to external tools through Actions if you wire up an API. That makes it a half-step toward automation, but a human still has to open it and start the conversation.

A real agent is different in kind, not degree. It runs on a trigger that isn’t you typing. It takes actions in other systems (sends the email, updates the CRM, files the record), it loops on its own output, and it keeps running when you close the laptop.

The three variables that decide it

When I’m scoping this for a client, I don’t ask how advanced they want to be. I ask three boring questions.

Memory. Does this need to remember a fixed body of knowledge (your brand voice, your pricing, your SOPs) and apply it consistently? A Project handles that beautifully. Upload the files once, and every conversation inherits them. You do not need an agent to get memory.

Repeatability. Is this the same task over and over with no judgment call in the middle, or is each run a little different? If a human’s taste is in the loop every time, keep the human and give them a Project. If the steps are identical and boring, that is a job for an agent.

Who else touches it. This is the variable people forget. A Project lives in your account, in your taste. The moment three teammates need to run the same process and get the same output, you need something with shared rules and guardrails, which pushes you toward a Custom GPT (shareable) or a real agent (no human needed at all).

Other people touch it →

Custom GPT / shared Project

• Team needs same output
• Still has judgment calls
• e.g. SDR first-draft replies

Real agent

• Runs without a human
• Identical steps every time
• e.g. nightly report build

Claude Project

• Just you
• Your taste in the loop
• e.g. your proposal drafting

Scripted automation

• Solo, fully repetitive
• Often no LLM needed
• e.g. a Zap or n8n flow

Repeatability →

Most operator work I see lives in the bottom-left and top-left. Very little of it actually belongs in the agent quadrant, which is exactly the opposite of how people spend their budget.

What it actually costs to reach for the wrong one

Here is the proposal-drafting job from the top, run three ways. Same task, wildly different total cost, because cost is not just the build, it is what you carry afterward.

Custom-built 'agent’ (the mistake)$9k+

Custom GPT with Actions~$1.2k

Claude Project (what he needed)~$240

The $9,000 build wasn’t just expensive once. It is now a thing that breaks, needs hosting, and needs the same developer every time the pricing sheet changes. The Project version updates by dragging a new file into the sidebar. The maintenance line is where the real money hides, and it is almost always invisible at the moment you say yes to the agent.

I built CentraOps doing the actual agent work, the kind that earns its keep. So I am not anti-agent. I am anti paying agent prices for chat-with-context jobs.

Your stack is the medication. The dose is whether a human still has to press go.

When an agent is genuinely the right call

To be fair to the other side: there are jobs where a Project is the wrong tool and you are leaving real value on the floor by not building the agent.

The tell is a trigger that isn’t you. A new lead hits the form at 2 a.m. and needs a reply before a human wakes up. An invoice lands in the inbox and needs to be coded and filed the same way every time. A weekly report needs to be assembled from four systems on Monday at 6 a.m. whether or not you remembered.

Those share the agent signature: event-driven, identical steps, no taste required, and the cost of a human doing it is real. That is the quadrant where the $9k is cheap. The mistake my client made wasn’t building an agent. It was building one for a job that had his judgment in the middle of every single run.

I’ll hold one honest caveat: this line moves. Agents are getting cheaper to build and easier to maintain every quarter, and the “afternoon in a Project” threshold will keep rising into territory that needs a real build today. So re-ask the three questions every few months, not once.

Do this before you commission anything

Before you pay anyone to build “an agent,” run the task through the three variables yourself. Most of the time you’ll discover you needed a Project and a quiet afternoon.

The serious move is rarely the most expensive one. It is matching the tool to the trigger, and most of the time the trigger is still you.